A Stream Protocol Implementation for an SCI-Based Cluster of Workstations
Jørgen S. Hansen, Povl T. Koch, Eric Jul · 1999
New high performance memory mapped networks such as SCI provide both low latency and high bandwidth. Many network servers can benefit from this performance, but use standard protocols such as TCP/IP which usually carry a high software overhead. We have implemented a prototype called SciStream, which exploits the low level performance of SCI cluster hardware, but retains compatibility with legacy protocols. The SciStream protocol is implemented at kernel level and is integrated with the TCP/IP subsystem of a general operating system. The implementation combines a number of optimizations and provides traditional protection and the ability to handle many concurrent processes. We present early performance measurements that show end-to-end latencies around 8.4 microseconds and bandwidths up to 72 Mbytes/s. 1 Introduction High performance network servers can no longer be implemented using a single computer---their service requirements are substantial---and growing. We base high performance...